golang-observability

$2c

• Use structured logging with log/slog — production services MUST emit structured logs (JSON), not freeform strings

• Choose the right log level — Debug for development, Info for normal operations, Warn for degraded states, Error for failures requiring attention

• Log with context — use slog.InfoContext(ctx, ...) to correlate logs with traces

• Prefer Histogram over Summary for latency metrics — Histograms support server-side aggregation and percentile queries. Every HTTP endpoint MUST have latency and error rate metrics.

• Keep label cardinality low in Prometheus — NEVER use unbounded values (user IDs, full URLs) as label values

• Track percentiles (P50, P90, P99, P99.9) using Histograms + histogram_quantile() in PromQL

• Set up OpenTelemetry tracing on new projects — configure the TracerProvider early, then add spans everywhere

• Add spans to every meaningful operation — service methods, DB queries, external API calls, message queue operations

• Propagate context everywhere — context is the vehicle that carries trace_id, span_id, and deadlines across service boundaries

• Enable profiling via environment variables — toggle pprof and continuous profiling on/off without redeploying

• Correlate signals — inject trace_id into logs, use exemplars to link metrics to traces

• A feature is not done until it is observable — declare metrics, add proper logging, create spans

• Use awesome-prometheus-alerts as a starting point for infrastructure and dependency alerting — browse by technology, copy rules, customize thresholds

Cross-References

See samber/cc-skills-golang@golang-error-handling skill for the single handling rule. See samber/cc-skills-golang@golang-troubleshooting skill for using observability signals to diagnose production issues. See samber/cc-skills-golang@golang-security skill for protecting pprof endpoints and avoiding PII in logs. See samber/cc-skills-golang@golang-context skill for propagating trace context across service boundaries. See samber/cc-skills@promql-cli skill for querying and exploring PromQL expressions against Prometheus from the CLI.

Go 1.26+: slog multi-handler

For simple fan-out to multiple slog handlers, prefer stdlib slog.NewMultiHandler before adding third-party handler-composition dependencies.

logger := slog.New(slog.NewMultiHandler(

    slog.NewJSONHandler(os.Stdout, nil),

    auditHandler,

))

Use third-party slog handler libraries only when the stdlib handler composition is insufficient.

The Five Signals

Signal

Question it answers

Tool

When to use

Logs

What happened?

log/slog

Discrete events, errors, audit trails

Metrics

How much / how fast?

Prometheus client

Aggregated measurements, alerting, SLOs

Traces

Where did time go?

OpenTelemetry

Request flow across services, latency breakdown

Profiles

Why is it slow / using memory?

pprof, Pyroscope

CPU hotspots, memory leaks, lock contention

RUM

How do users experience it?

PostHog, Segment

Product analytics, funnels, session replay

Detailed Guides

Each signal has a dedicated guide with full code examples, configuration patterns, and cost analysis:

-

Structured Logging — Why structured logging matters for log aggregation at scale. Covers log/slog setup, log levels (Debug/Info/Warn/Error) and when to use each, request correlation with trace IDs, context propagation with slog.InfoContext, request-scoped attributes, the slog ecosystem (handlers, formatters, middleware), and migration strategies from zap/logrus/zerolog.

-

Metrics Collection — Prometheus client setup and the four metric types (Counter for rate-of-change, Gauge for snapshots, Histogram for latency aggregation). Deep dive: why Histograms beat Summaries (server-side aggregation, supports histogram_quantile PromQL), naming conventions, the PromQL-as-comments convention (write queries above metric declarations for discoverability), production-grade PromQL examples, multi-window SLO burn rate alerting, and the high-cardinality label problem (why unbounded values like user IDs destroy performance).

-

Distributed Tracing — When and how to use OpenTelemetry SDK to trace request flows across services. Covers spans (creating, attributes, status recording), otelhttp middleware for HTTP instrumentation, error recording with span.RecordError(), trace sampling (why you can't collect everything at scale), propagating trace context across service boundaries, and cost optimization.

-

Profiling — On-demand profiling with pprof (CPU, heap, goroutine, mutex, block profiles) — how to enable it in production, secure it with auth, and toggle via environment variables without redeploying. Continuous profiling with Pyroscope for always-on performance visibility. Cost implications of each profiling type and mitigation strategies.

-

Real User Monitoring — Understanding how users actually experience your service. Covers product analytics (event tracking, funnels), Customer Data Platform integration, and critical compliance: GDPR/CCPA consent checks, data subject rights (user deletion endpoints), and privacy checklist for tracking. Server-side event tracking (PostHog, Segment) and identity key best practices.

-

Alerting — Proactive problem detection. Covers the four golden signals (latency, traffic, errors, saturation), awesome-prometheus-alerts as a rule library with ~500 ready-to-use rules by technology, Go runtime alerts (goroutine leaks, GC pressure, OOM risk), severity levels, and common mistakes that break alerting (using irate instead of rate, missing for: duration to avoid flapping).

-

Grafana Dashboards — Prebuilt dashboards for Go runtime monitoring (heap allocation, GC pause frequency, goroutine count, CPU). Explains the standard dashboards to install, how to customize them for your service, and when each dashboard answers a different operational question.

Correlating Signals

Signals are most powerful when connected. A trace_id in your logs lets you jump from a log line to the full request trace. An exemplar on a metric links a latency spike to the exact trace that caused it.

Logs + Traces: otelslog bridge

import "go.opentelemetry.io/contrib/bridges/otelslog"

// Create a logger that automatically injects trace_id and span_id

logger := otelslog.NewHandler("my-service")

slog.SetDefault(slog.New(logger))

// Now every slog call with context includes trace correlation

slog.InfoContext(ctx, "order created", "order_id", orderID)

// Output includes: {"trace_id":"abc123", "span_id":"def456", "msg":"order created", ...}

Metrics + Traces: Exemplars

// When recording a histogram observation, attach the trace_id as an exemplar

// so you can jump from a P99 spike directly to the offending trace

obs := histogram.WithLabelValues("POST", "/orders")

if eo, ok := obs.(prometheus.ExemplarObserver); ok {

    eo.ObserveWithExemplar(duration, prometheus.Labels{"trace_id": traceID})

} else {

    obs.Observe(duration)

}

Migrating Legacy Loggers

If the project currently uses zap, logrus, or zerolog, migrate to log/slog. It is the standard library logger since Go 1.21, has a stable API, and the ecosystem has consolidated around it. Continuing with third-party loggers means maintaining an extra dependency for no benefit.

Migration strategy:

• Add slog as the new logger with slog.SetDefault()

• Use bridge handlers during migration to route slog output through the existing logger: samber/slog-zap, samber/slog-logrus, samber/slog-zerolog

• Gradually replace all zap.L().Info(...) / logrus.Info(...) / log.Info().Msg(...) calls with slog.Info(...)

• Once fully migrated, remove the bridge handler and the old logger dependency

Definition of Done for Observability

A feature is not production-ready until it is observable. Before marking a feature as done, verify:

• Metrics declared — counters for operations/errors, histograms for latencies, gauges for saturation. Each metric var has PromQL queries and alert rules as comments above its declaration.

• Logging is proper — structured key-value pairs with slog, context variants used (slog.InfoContext), no PII in logs, errors MUST be either logged OR returned (NEVER both).

• Spans created — every service method, DB query, and external API call has a span with relevant attributes, errors recorded with span.RecordError().

• Dashboards and alerts exist — the PromQL from your metric comments is wired into Grafana dashboards and Prometheus alerting rules. Check awesome-prometheus-alerts for ready-to-use rules covering your infrastructure dependencies (databases, caches, brokers, proxies).

• RUM events tracked — key business events tracked server-side (PostHog/Segment), identity key is user_id (not email), consent checked before tracking.

Common Mistakes

// ✗ Bad — log AND return (error gets logged multiple times up the chain)

if err != nil {

    slog.Error("query failed", "error", err)

    return fmt.Errorf("query: %w", err)

}

// ✓ Good — return with context, log once at the top level

if err != nil {

    return fmt.Errorf("querying users: %w", err)

}

// ✗ Bad — high-cardinality label (unbounded user IDs)

httpRequests.WithLabelValues(r.Method, r.URL.Path, userID).Inc()

// ✓ Good — bounded label values only

httpRequests.WithLabelValues(r.Method, routePattern).Inc()

// ✗ Bad — not passing context (breaks trace propagation)

result, err := db.Query("SELECT ...")

// ✓ Good — context flows through, trace continues

result, err := db.QueryContext(ctx, "SELECT ...")

// ✗ Bad — using Summary for latency (can't aggregate across instances)

prometheus.NewSummary(prometheus.SummaryOpts{

    Name:       "http_request_duration_seconds",

    Objectives: map[float64]float64{0.99: 0.001},

})

// ✓ Good — use Histogram (aggregatable, supports histogram_quantile)

prometheus.NewHistogram(prometheus.HistogramOpts{

    Name:    "http_request_duration_seconds",

    Buckets: prometheus.DefBuckets,

})